- Sources: primary
- Summary: The preprint reports that coding assistants opened a supply-chain provenance signal in 9 of 1,920 install trials, a rate of 0.5 percent, and that no trial ran a verification command. The authors deposited the protocol, the seed and the analysis plan with a DOI before the first trial ran, and scored behaviour from container logs rather than from what the assistant reported doing. The run logs and a cost ledger are released with the paper.
- Why it matters: Price did not buy verification, because the assistant that opened a provenance signal most often cost $0.10 per trial while the most capable one, at $1.00, verified nothing.
- Follow-up: Whether the result survives peer review and whether any assistant vendor reports a change in this behaviour.
send feedback on this story